Progressive Multi-level Distillation for Domain Adaptive Object Detection

Domain adaptive object detection (DAOD) is intrinsically a multi-layered challenge. While existing adversarial alignment or self-training methods offer partial solutions, they often struggle with training instability or the accumulation of semantic noise. In this paper, we propose a Progressive Multi-level Distillation (PMD) framework, which systematically mitigates the domain shift via a “Structural-to-Spectral-to-Semantic” refinement pipeline. Our core philosophy is to rectify the cross-domain representation at three increasing levels of abstraction: Hierarchical Feature Alignment (HFA) for multi-scale structures; Tensor Low-rank Distillation (TLD) using SVD to purify latent manifolds; and CLIP-Guided Pseudo-Label Calibration Module (CPCM) for semantic correction and prevention of pseudo-label error accumulation. These three components form a unified pipeline that systematically refines feature representations from low-level structural alignment to high-level semantic calibration, thereby enhancing overall adaptability. Extensive experiments conducted across three representative cross-domain scenarios demonstrate that our proposed framework achieves superior performance over existing state-of-the-art methods, validating the effectiveness of progressive multi-level distillation.